My AI Coworker Maintains My Smart Home, Part 1: The Audit

My Home Assistant had 1,583 automations. I could name maybe forty. This series is about the AI agent that now co-manages the house with me, and it starts where we started: an audit that found 88% of my home’s daily activity was one runaway automation checking email every 30 seconds.

the state of the house

Confession by census. When my AI agent first pulled the automation registry, the numbers came back: 1,583 automation entities. 629 disabled. 688 that had never fired, not once. 817 enabled but stale, silent for six months or more. Actual living automations, fired in the last week: 55.

Growth explains most of it, not sloth. Lock-management software generates hundreds of per-slot automations by design; integrations come and go and leave orphans; a rebuilt automation leaves its ancestor behind, disabled, forever. Fifteen years in any codebase produces the same sediment. Except a codebase gets code review, and my house had never had one.

enter the coworker

The agent (his name is Loby; he lives in my notes vault and works on a schedule) got the job no human gives themselves: read the whole registry, trace what fires, write down what you find. Nightly and weekly scheduled runs against the HA REST and WebSocket APIs, findings written to markdown in the vault where I read them with coffee.

The first fires-per-day report had a headline. One automation, 2,874 fires in 24 hours, 88% of all activity in the house. The entire rest of the home, every light, lock, blind, camera, and sensor reaction, split the remaining 12%. The culprit checked my email for Pokémon preorder announcements every 30 seconds (that fiasco got its own post last week). The kicker: every downstream step in it was disabled. The house’s most active automation, by nine to one, did nothing at all.

I laughed, then I read the rest of the report and stopped. A leak-sensor alert, the kind that protects the water heater, notified two phones that hadn’t existed in years. The automation was enabled and healthy; delivery went nowhere. Silent failure in the direction that matters. The audit caught it because the agent checks that notify targets exist, a thing no human re-verifies after replacing a phone.

what a house code review looks like

The weekly review reads like a colleague’s PR comments, quotes from actual findings:

Volume outliers: sort by fires-per-day, question the top five. An AI forecast blurb regenerating every 15 minutes, for a dashboard nobody looks at between mornings, got dialed to hourly.

Dead targets: automations pointing at entities that no longer exist. Renamed devices are the usual killer; my morning briefing died silent for five months this way.

Graveyard candidates: hundreds of generated lock automations for locks replaced years ago. Deleting them plus flushing ghost states took the registry from 1,583 to 966 in one cleanup pass. A quarter of my “smart home” was headstones.

Cosmetic rot, which is future functional rot: an automation named _offf, a blank one named new_automation, names that describe triggers the trigger no longer matches.

the part I didn’t expect

Config-to-git came out of the same effort: nightly, the agent syncs my HA config into a git repo, secrets excluded. The audit found problems, and git answers the question after the finding: when did this change, and what else changed that day? An automation misbehaving on Tuesday plus a diff from Monday night is a five-minute diagnosis. Before git, the answer lived in whatever I remembered about my own weekend tinkering, an unreliable narrator on his best day.

Numbers so far: 1,583 automations at the start, 966 after cleanup. One runaway killed (2,880 daily runs → 1). One broken safety alert fixed before water found it. A backup system that had, per its own event log, never completed a successful automatic backup: caught, reported, repaired.

why an agent instead of a dashboard

Dashboards report what you ask; my failures lived in questions I didn’t know to ask. “Which notify targets are dead” was not on any dashboard I would have built. The agent’s job description is closer to go looking for trouble and write it down, and writing findings into the same vault I keep everything else in means review costs me nothing extra. The report is sitting next to my morning notes, already prioritized, when I sit down.

A maintainer that never gets bored, checking work I did years ago against the house as it exists today. That arrangement is the actual upgrade, more than any single fix it produced.

the series

Part 2 covers the morning briefing: how the agent’s overnight notes become a 60-second spoken digest, and the data-injection rule that keeps it truthful. Part 3 covers the no-nag list, the boring-sounding file that fixed alert fatigue in my house: a human-editable table of known-broken things that every report honors, so the house nags about new problems only. One table row silences a broken blind everywhere; deleting the row un-silences it.

A smart home accumulates automation the way a garage accumulates cables. You can sort it yourself once a decade, or you can hire a coworker who finds cables fun. Mine works nights.